NeuroImage 120 (2015) 428–440

Contents lists available at ScienceDirect

NeuroImage

j ourna l homepage: www.e lsev ie r .com/ locate /yn img

Adaptation of the human visual system to the statistics of letters andline configurations

Claire H.C. Chang a,⁎, Christophe Pallier b,c,d,e, Denise H. Wu f, Kimihiro Nakamura b,c,e,g, Antoinette Jobert b,c,e,W.-J. Kuo a, Stanislas Dehaene b,c,e,g

a Institute of neuroscience, National Yang-Ming University, Taipei, Taiwanb Cognitive Neuroimaging Unit, INSERM, Gif-sur-Yvette, Francec CEA, DSV, I2BM, NeuroSpin Center, Paris, Franced Centre National de la Recherche Scientifique, Paris, Francee University Paris-Sud, Paris, Francef National Central University, Jhongli, Taiwang Collège de France, Paris, France

⁎ Corresponding author at: Room 812, 8F, Library, InforNational Yang-Ming University, No. 155, Sec. 2, Linong Str

E-mail address: hcchang73@gmail.com (C.H.C. Chang)

http://dx.doi.org/10.1016/j.neuroimage.2015.07.0281053-8119/© 2015 Elsevier Inc. All rights reserved.

a b s t r a c t

a r t i c l e i n f o

Article history:Received 5 May 2015Accepted 8 July 2015Available online 17 July 2015

Keywords:Environmental statisticsAdaptationVisual recognitionReadingLiteracyfMRI

By adulthood, literate humans have been exposed to millions of visual scenes and pages of text. Does the humanvisual systembecome attuned to the statistics of its inputs? Using functionalmagnetic resonance imaging,we ex-amined whether the brain responses to line configurations are proportional to their natural-scene frequency. Tofurther distinguish prior cortical competence from adaptation induced by learning to read, we manipulatedwhether the selected configurations formed letters andwhether theywere presented on the horizontalmeridian,the familiar location where words usually appear, or on the vertical meridian. While no natural-scene frequencyeffectwas observed,weobserved letter-status and letter frequency effects on bilateral occipital activation,mainlyfor horizontal stimuli. The findings suggest a reorganization of the visual pathway resulting from reading acqui-sition under genetic and connectional constraints. Even early retinotopic areas showed a stronger response toletters than to rotated versions of the same shapes, suggesting an early visual tuning to large visual featuressuch as letters.

© 2015 Elsevier Inc. All rights reserved.

Introduction

Many neuroscientists and theorists have proposed the idea that thevisual system has internalized the statistical properties of the environ-ment (Berkes et al., 2011; Geisler, 2008; Girshick et al., 2011; Longand Purves, 2003; Shepard, 2002; Simoncelli and Olshausen, 2001).For example, environmental statistics have been proposed to be thebasis of the Gestalt rules of proximity (Brunswik, 1956) and the princi-ple of good continuation (Gilbert et al., 2001b). The adaptation of thevisual system to environmental regularities could occur both at the evo-lutionary scale (Shepard, 2002) and during ontogenetic development(Berkes et al., 2011; Blakemore and Cooper, 1970; Held and Hein,1963). In the Bayesian perspective, environmental statistics get inter-nalized and later enter as a prior which is used to help disambiguate fu-ture inputs (Kersten et al., 2004; Knill and Pouget, 2004). Classical visualillusions such as the horizontal-vertical illusion (greater apparent size ofa vertical bar compared to a horizontal bar) may be explained by scenestatistics (Howe and Purves, 2002). This and other illusions may be

mation, and Research Building,eet, Taipei, 112 Taiwan ROC..

accounted for by supposing that early visual neuronal circuits are mod-ified by experience, such that greater populations of cells are assigned tomore frequent features of the environment (Girshick et al., 2011) andthat their horizontal connections internalize the statistics of featureco-occurrence (Hess et al., 2003).

In the present study, we examined whether the frequency distribu-tion of line configurations in the environment is reflected in the humanvisual system. Changizi et al. (2006) discovered an interesting statisticalregularity in the frequency with which the topological configurationsformed by image contours, such as T, L, or X configurations, occur inthe visual environment. They counted the frequency of each topologicalconfiguration of two or three contour lines in pictures of the natural orartificial human environment, and observed a systematic ordering(Fig. 1a). For instance, amongst the two-line configurations, the “L” con-figuration was always more frequent than “T”, which in turn was morefrequent than “X”. Crucially, this is not the case in simple random ar-rangements of lines. Furthermore, human visual signs, as taken from al-phabets, logographic writing systems and other symbol systems,followed the same statistical distribution, such that the frequencyranks of the configurations in these two domains were positively corre-lated (Fig. 1a). In other words, the frequency distribution of line config-urations in human cultural signsmimicked that found in natural scenes.

Fig. 1. Stimulus design for Experiment 1. A: Correlation between the frequency of simple line configurations in natural scenes and in writing systems (redraw from Changizi et al., 2006).The x and y axes indicate the rank of each configuration according to the corresponding frequency. Configurations drawn in dark gray have the lowest ranking in both domains and theircoordinateswere shifted slightly to allow thedisplay of thewhole configurations. B: Examples of stimuli used in fMRI. 15 line configuration typeswere selected. 20 stimuli of the same type(flashed for 200ms, separated by 200ms blanks) were presented in short blocks of 8 seconds, separated by 6–8 s resting periods. The subject’s task was to respond to the single-line con-figuration (top left) which appeared occasionally inside the blocks.

429C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

Such a strong link between environmental statistics and cultural inven-tions is in agreementwith the "neuronal recycling hypothesis"wherebynovel cultural acquisition such as writing maps onto pre-existing corti-cal systems, thus constraining the range of cross-cultural variations(Dehaene and Cohen, 2011). According to this hypothesis, Changiziet al.’s (2006) finding implies that configurations that are frequently ob-served in the natural environment aremore likely to be selected as visu-al signs, because they are better encoded in the visual system (Dehaene,2009).

The above argument is based solely on statistical evidence, and lacksdirect evidence that line configuration statistics are encoded in the visu-al system.Neurophysiological evidence does suggest that neurons in theprimate infero-temporal cortex can be sensitive to specific line configu-rations that form non-accidental topological properties (Brincat andConnor, 2004, 2006; Tanaka, 2003). However, these studies have notyet investigated whether the cortical representation of these featuresmimics their distribution in natural scenes. Here, we used fMRI inhumans to investigate this issue. Our hypothesis was that visual activa-tion in response to line configurations should be directly proportional totheir natural-scene frequency.

Where in the visual pathway might this effect occur? A predictionfor the locus of the natural-scene frequency effect could be madebased on hierarchical models of visual recognition (Dehaene et al.,2005; DiCarlo et al., 2012; Rolls and Stringer, 2006; Serre et al., 2007;Ullman, 2007). All of these models assume that the ventral occipito-temporal pathway comprises a hierarchy of neural detectors withprogressively larger receptivefields, each tuned to increasingly complexand abstract combinations of visual features. In humans, the Local Com-bination Detectors model (Dehaene et al., 2005) assumes that writtenword recognition rests on a reorientation of this architecture towardsthe detection of letters and their combinations. Based on several priorfMRI experiments (Dehaene et al., 2004; Vinckier et al., 2007), themodel proposes specific cortical areas for each step: line configurationsand letter fragments in area V2 and V4, abstract letter identities andtheir combinations in the more anterior visual word form area (VWFA)(Cohen et al., 2002). Under the neuronal recycling hypothesis, evenprior to reading acquisition, these areas may already exhibit a bias forrecognizing line configurations, which would make it particularly suit-able for recognition of written words (Dehaene et al., 2005; Dehaene,2009; Szwed et al., 2011).

In addition to V2, V4 and VWFA, one should also consider the possi-bility that the primary visual cortex itself may exhibit sensitivity, not

only to elementary contours, but also to their frequent combinations.Recent electrophysiological (McManus et al., 2011) and imaging(Sigman et al., 2005) studies have revealed that training in shape detec-tion changes cortical responses even in the calcarine cortex, indicatingthat experience could induce a sensitivity to complex visual featuresin early retinotopic areas V1 and/or V2. fMRI studies of reading indicatethat even area V1 is more activated by letter strings than by scrambledstimuli with matched visual features (Szwed et al., 2011, 2014). Indeed,the calcarine cortex, at the location of area V1, shows a strongerresponse to horizontal checkerboards in literate, who used to readhorizontally, than in illiterate subjects (Dehaene et al., 2010). Thosefindings suggest that reading acquisition may lead to perceptual learn-ing for frequent letter shapes in area as early as V1. In this case, an effectof the natural-scene frequencies of line configurations might also beobserved in early retinotopic cortex.

In summary, we aim to test whether the frequency distribution ofsimple line configurations in natural scenes is reflected in the visualcortex. In Experiment 1, we study the fMRI responses to the line config-urations studied by Changizi et al. (2006). Given that themost frequentconfigurations in natural scenes are also those most frequently used inhuman writing systems, the existence of such effect may support theview that humanwriting systems have evolved fromprior cortical com-petence. Furthermore, in literate adults, the adaptation to environmen-tal statistics includes a novel cultural environment: written texts. Thus,onemight expect the processing of simple line configurations to also beunder the influence of reading experience, a prediction which is furtherinvestigated in Experiment 2.

Experiment 1

In Experiment 1, we collected fMRI data in 18 subjects while theysimply viewed arrays comprising 15 different types of line configura-tions, selected to span a broad range of natural-scene frequencies, ascomputed by Changizi et al. (2006)(Fig. 1). We used a correlationapproach to probe the whole brain for activations correlated with thelogarithm of natural scene frequencies.

Methods

ParticipantsEighteen right-handed (9 female), 18–30 year-old native French

speakers, participated in the present fMRI experiment. They had no

430 C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

history of neurological or psychiatric disease, and had normal orcorrected-to-normal vision. Written informed consents were given.The project was approved by the local ethics committee.

StimuliFifteen configurations were selected from the paper by Changizi

et al. (2006), which provides the frequencies of line configurations inpictures of the human environment (landscapes, cities, etc.). In thisstudy, we use as a short-hand the expression ‘natural-scene frequency’to refer to the logarithm of the average frequencies of line configura-tions in such pictures.

For each type of line configuration, we used a matlab program togenerate 10 images, each comprising 35 randomly oriented items ofthe same type (Fig. 1b). The itemwidth was 0.83-1.22 degrees of visualangle. The images were 18.9 by 18.9 degrees of visual angle. Totalcontour length and number of “on” pixels were matched (the standarddeviation of the numbers of “on” pixels was less than 0.1% across allconditions).

Design and procedureEach participant took part in six fMRI runs. The total scanning time

was around 42minutes. Each run lasted about 7 minutes and contained30mini-blocks of 8 s separated by rest periods of 4–8 s. Eachmini-blockcomprised 20 images of the same type of line configuration, each ofwhich was presented for 200 ms after a fixation interval of 200 ms.Each run comprised two mini-blocks of each of the 15 distinct types ofline configuration. The 30 blocks were ordered randomly. To maintainthe participants’ attention on the visual stimuli with a minimally de-manding task, participants were required to monitor the stimulusstream for the presence of a target probe consisting of a picture withsingle-line bars, also displayed for 200 ms. The target appeared in halfof the mini-blocks, and participants were instructed to press a buttonas fast as possible upon seeing it. Blocks with a target were randomlychosen. The target always occurred in the middle of blocks (replacingone of the images 6–12 within the block of 20 images). In an effort tomaintain attention throughout each 8-second block, subjects were nottold that blocks could only contain at most one target.

MRI acquisitionThe acquisition was performed with a 3-Tesla Siemens Tim Trio

system. One anatomical image (voxel = 1x1x1.1 mm) and a total of1092 functional images were acquired using an Echo-Planar sequencesensitized to the BOLD effect (TR = 2.4 secs, TE = 30 msecs, Matrix =64x64; Voxel size = 3x3x3 mm; 40 slices in ascending order).

Data analysisData processing was performed with SPM8 (Wellcome Department

of Cognitive Neurology, software available at http://www.fil.ion.ucl.ac.uk/spm). The anatomical scan was spatially normalized to the avg152T1-weighted brain template defined by the Montreal NeurologicalInstitute using the default parameters (nonlinear transformation).Functional volumes were realigned to correct for movements, spatiallynormalized using the parameters obtained from the normalizationof the anatomy, and smoothed with an isotropic Gaussian kernel(FWHM = 5 mm).

In a first SPM model, experimental effects at each voxel were esti-mated using a multi-run design matrix modeling the 15 configurations,the probe trials, and the 6 movement parameters computed at the re-alignment stage. Each blockwasmodeled as an epoch lasting 8 seconds,and each probe trial as a punctual event. The regressors were created byconvolving these epochs by the standard SPM hemodynamic responsefunction. Contrasts averaging the regression weights associated witheach configuration were computed.

These estimates of individual effect sizes were entered in a second-level analysis with one regressor for each configuration and each partic-ipant (one-way within-subject ANOVA model). To search for regions

showing an effect of natural-scene frequency, we used a contrast withweights proportional to log natural-scene frequency, testing for increas-ing activation across the configurations in the ANOVA model. We alsotested second-level regression models pitting two variables againsteach other, as described further below. Unless otherwise stated, statis-tics were thresholded at voxel wise p b 0.001 (uncorrected), with anadditional correction for multiple comparisons across the whole-brainvolume based on cluster extent (p b 0.05, FDR corrected). Regionsshowing significant effects were labeled with an automated anatomicallabeling system (AAL; Tzourio-Mazoyer et al., 2002).

Results

Behavioral resultsReaction times (RT) outside the range of individual mean ± 3 sd

were excluded. Across participants, the mean RT was 445 ms (SE =13 ms, range = 348–633 ms), and the mean accuracy was 97 %(SE = 1 %, range = 82–100%). The RTs and accuracies of each configu-ration were listed in Appendix Table 1. Repeated one-way ANOVAs re-vealed small but significant differences between configurations in RT(F (14, 238) = 2.73; p b .01) and accuracies (F(14, 238) = 2.15;p b .05). However, natural-scene frequency was not significantly corre-lated with either RT (r = 0.32, p = .24) or accuracy (r = −0.24, p =.39). The behavioral results confirmed that the participants maintainedtheir attention on the visual presentation.

Imaging resultsWhole-brain analysis revealed a bilateral occipital cluster with a sig-

nificant positive correlation indicating increasingly stronger activationfor configurations with increasingly higher natural-scene frequency inearly retinotopic areas (Fig. 2 and Table 1).

The scatter plots in Fig. 2 illustrate how occipital activation variesacross the 15 line configurations. Although there is a clear trend as afunction of natural-scene frequency, somedispersion in activation is ap-parent. Furthermore, one may observe that configurations correspond-ing to letters (shown in red), which are all of high natural-scenefrequency, yield stronger activations than other configurations withnearly-equivalent frequency. To formally assess the effect of letterstatus, we created a multiple regression model with one regressor perparticipant and two regressors of interest: natural-scene frequencyand letter versus non-letter status (X, T, L, H, Y and F configurationswere counted as letters, although note that they often appeared asrotated in the display; this factor will be controlled in Experiment 2).In this model, the effect of natural-scene frequency ceased to reachsignificance anywhere in the brain, and instead there was a significanteffect of letter status in bilateral occipital cortex, including bilateral V1,V2, and left V3 (Fig. 3 and Table 2).We also tested the natural-scene fre-quency effect within only the non-letters, again without any significantresults. Thus, the results suggest that letter status, rather than frequency,drives occipital fMRI activation in educated human adults.

In an effort to confirm this conclusion while controlling for other vi-sual variables thatmay be confoundedwith letter/non-letter status and/or natural-scene frequency, wemeasured several parameters of the dis-plays: convex area (the surface of the smallest convex polygon that con-tains a single line configuration item), number of line junctions, numberof strokes (2 or 3), number of angles, and number of terminals (endingpoints of a line). The values of the variables for each configuration typeare provided in Appendix Table 1. We also added as a potential con-found the average response time for target detection in the correspond-ing block. We created several regression models in which each of thesevariables was pitted against natural-scene frequency, and observed thatin the models including letter status or convex area, the variable 'natu-ral-scene frequency’ no longer yielded a significant effect. In a modelwith letter status and convex area, we only found higher occipital acti-vation for letters than non-letters (Fig. 3). We further tested this letterstatus effect in models systematically including letter status and one

Fig. 2. Brain regions showing a significant increase in activation with natural-scene frequency (N= 18, thresholded at T N 3.12, voxelwise p b .001, uncorrected; clusterwise p b 0.05, FDRcorrected). The scatterplots show the average fMRI activation for all 15 configurations in left and right occipital peaks. Error bars represent 1 standard error across participants after sub-traction of each participant’s individual mean. Letter-like configurations are displayed in red.

L

431C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

of the other variables as regressors of interest. Letter status always sur-vived as the dominant determinant of occipital activation (Fig. 4).

We also tested for a reversed letter-status effect, namely, larger acti-vations for non-letters than letters, as well as the negative correlationbetween brain activations and natural-scene frequency. No such effectswere found in either the ANOVAmodel or the regression model includ-ing letter status and natural-scene frequency.

Discussion

In Experiment 1,we tested the hypothesis that the natural-scene fre-quencies of line configurations are reflected in the human visual system.As predicted, a positive correlation between natural-scene frequencyand brain activation was found only in bilateral occipital visual areas,at an anatomical location corresponding to area V1/V2 and a smallpart of left V3. However, we also found that this effect could be drivenby a partially confounded variable, namely, whether or not a givenline configuration forms a letter of the Roman alphabet. As observedby Changizi et al. (2006), in all cultures, the shapes that are used asletters tend to be of high natural-scene frequency. Nevertheless, ourstimulus set included some non-letter line configurations with anatural-scene frequency nearly as high as that of the letters. Multipleregression analyses suggested that letter status, not natural-scene fre-quency, was responsible for the changes in occipital activation.

Such an effect of letter status is compatible with prior observationsthat early visual cortex is modified by literacy acquisition (Dehaeneet al., 2010) and becomes sensitive to letters strings more than toother stimuli of matched complexity (Szwed et al., 2011, 2014). Notethat this effect is not incompatible with the general hypothesis thatthe visual system internalizes the statistics of environmental inputs. Itshould be acknowledged that, for highly literate subjects, the environ-ment most likely includes a high proportion of text, which may there-fore bias the statistics away from those of natural scenes and towardsthose of the subject’s writing system.

Experiment 1, however,wasnot specifically designed to test for a let-ter effect, but solely to investigate the effect of natural-scene frequency,and the letter effect was only seen in a post-hoc analysis. In Experiment

Table 1Brain regions showing natural-scene frequency effect in Experiment 1.

Cluster Size T X Y Z

659 Occipital Calcarine R 6.32 15 −101 0Cuneus R 6.06 18 −97 7Inf. L 5.75 −12 100 −8

2, we therefore aimed to provide a replication in which the effects ofnatural-scene frequency and letter status were manipulated indepen-dently. To this aim, we capitalized on the fact that, in written texts,letters appear at a specific angle. Beyond about 45 degrees of rotations,the recognition of letters andwords becomes severely degraded, accom-panied by a sudden onset of parietal lobe activations suggesting serial

Fig. 3.Disappearance of the effect of natural-scene frequency once other variables are con-sidered. The results of two regressionmodels are shown, each containing two regressors ofinterest: the line configuration frequency in natural scenes and either the letter status(whether the configuration forms a letter of the Roman alphabet or not) or the convexarea (estimating the surface occupied by an individual line configuration item in the dis-play). Images are SPMt maps (N= 18, thresholded at T N 3.12, voxelwise p b .001, uncor-rected; clusterwise p b 0.05, FDR corrected).

Table 2Brain regions showing letter status effect in Experiment 1.

Cluster Size T X Y Z

531 Occipital Mid. L 9.11 −18 −94 −5Mid. L 5.04 −27 −85 10Mid. L 4.80 −39 −91 4

513 Occipital Mid. R 8.45 24 −91 7Calcrine R 8.05 18 −100 1Lingual R 4.58 6 −85 −11

432 C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

effortful deciphering (Cohen et al., 2008). In Experiment 2, we thereforepresented the very same topological line configuration at two differentangles, only one of which corresponded to a letter. We selected 8 letters(AKYHXFTL) and created a fixed set of 8 corresponding non-letter stim-uli by rotation or symmetry (Fig. 5). Although the range of variation innatural-scene frequency was smaller than that in Experiment 1, the 8configurations still spanned more than two orders of magnitude in theChangizi et al. scale (Appendix Table 2), thus achieving an orthogonaldesign with independent factors of letter status and natural-scenefrequency.

Experiment 2 also included another manipulation of the retinotopiclocation of the items. In Experiment 1, we attempted to maximize theeffect by covering the available visual field with many items. In Experi-ment 2, the stimuliwere presented in amore restricted part of the visualfield, either along the horizontal or the vertical meridian (Fig. 5). Be-cause the Roman alphabetic system is based on horizontal lines readfrom left to right, expert readers get considerablymore training in letterdecoding along the horizontal meridian. Although a page of textmay filla large part of the visual field, the reader’s attention is typically focusedon the letters left and right of fixation, and this is likely to have a deter-minant effect on the acquisition of visual expertise. Indeed, behavioraland brain-imaging evidence suggests an enhanced representation ofstimuli presented at or near the horizontal meridian in expert readers(Dehaene et al., 2010; Nazir et al., 2004). Accordingly, one may predicta larger effect of letter status in retinotopic cortical regions coding forthe horizontal meridian, than in those coding for the vertical meridian.

L

Fig. 4.Occipital activation is primarily determined by letter status, evenwhen other confoundedtwo regressors: letter status and one of the other variables (N = 18, thresholded at T N 3.12, v

Conversely, one may hope to find a purer effect of natural-scenefrequency, less strongly affected by reading experience, for stimulipresented along the vertical meridian.

Experiment 2

Methods

ParticipantsAfter exclusion of one subject (see below), 18 right-handed, 18–30

year-old native French speakers (10 female, 8 male) were retained inthis fMRI experiment. They had no history of neurological or psychiatricdisease and normal or corrected to normal vision. Written informedconsents were given. The project was approved by the local ethicscommittee.

StimuliEight letters were selected: AKYHXFTL. Wewrote a matlab program

to display them in simple line form. For each letter, we selected a trans-formation (flipping and/or rotation ranging from 55 to 180 degrees) tocreate a corresponding non-letter. We endeavored to match letters andnon-letters for the number of vertical and horizontal lines, with thesingle exception of configuration “X”. Thiswas done to avoid a confoundbetween letter/non-letter status and line orientation, since it is knownthat cellswhose receptive fields fall near the vertical and horizontalme-ridians exhibit a preference for vertical and horizontal lines, respectively(Furmanski and Engel, 2000).

Pictures corresponding to thirty-two conditions (8 configurations x2 letter status x 2 presentation orientation) were created (Fig. 5). Eachpicture contained 10 items of the same line configuration, eitherhorizontally or vertically aligned, with a small spatial jitter (Fig. 5).The size of each item was proportional to the distance from fixation,in order to compensate for the increase in receptive field size and thecorresponding loss in spatial resolution. The formula we used for itemsize (item size in degrees = 0.15 x distance from the fixation indegrees + 0.48) was derived from Harvey and Dumoulin (2011). Thepictures were 19.7 by 19.7 degrees of visual angle.

variables are taken into account. Each image is a SPMtmap from a regressionmodel withoxelwise p b .001, uncorrected; clusterwise p b 0.05, FDR corrected).

Horizontal Vertical

Letter

Non-letter

B ALetters

Non-letters

A K Y H X F T L

Fig. 5. Stimulus design for Experiment 2. The experimentwas designed to test the hypothesis that early visual cortices would be especially responsive to letters presented in their normalorientation and at the usual horizontal location. A: Eight line configurations corresponding to letter shapes were selected and were presented either in normal upright form (letter con-dition), or in an unusual rotated form (non-letter condition). B: Sample displays illustrating the 2 x 2 factorial design manipulating letter status (letter vs non-letter) and orientation ofpresentation (identical configurations were presented along the vertical or the horizontal meridian).

433C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

Design and procedureThe procedure was similar to Experiment 1, except the number of

blocks (32) and the task. In order to better equate task difficulty acrossthe different line configurations, the bar detection task used in Experi-ment 1 was replaced by a color detection task: we asked the participantto press a buttonwhenever they detected a colored item in the pictures.The colored probe appeared 40 times in a pseudo-randomized order, sothat the probe never occurred as the first or last trial in a block, and thatany two probes were separated by at least three trials.

MRI acquisitionThe acquisition was performed with a 3 Tesla Siemens Tim Trio sys-

tem. One anatomical image (voxel= 1×1×1.1mm) and a total of 1890functional imageswere acquired using amultiband sequence developedby the Center forMagnetic Resonance Research (CMRR) (Feinberg et al.,2010; Moeller et al., 2010; Xu et al., 2013) and sensitized to the BOLDeffect (TR = 1.5 secs, TE = 32 msecs, Matrix = 128×128; Voxelsize = 1.5×1.5×1.5 mm; 54 axial slices covering the occipital andmost inferior part of the temporal lobe).

MRI analysesThe data was preprocessed with the same procedure as in Experi-

ment 1. In the first level SPM models, experimental effects at eachvoxel were estimated using a multi-run design matrix modeling theeight configurations, targets in the four position (right, left, upper,lower), and the six movement parameters. Each block was modeled asan epoch lasting 8 seconds, and each probe trial as event with zeroduration. The regressors were created by convolving these epochs bythe standard SPMhemodynamic response function. Contrasts averagingthe regression weights associated with each configuration werecomputed.

These estimates of the individual effect sizes were entered in asecond-level analysis with one regressor for each configuration, aswell as each participant. The analysis was donewithin a mask includingthe occipital regions, lingual gyrus, and fusiform gyrus from the WakeForest University (WFU) PickAtlas (Maldjian et al., 2003). Specific re-gions of interest (ROIs) described in the next paragraph were also in-cluded. For the voxel-based analysis, the activations were thresholdedat p b .005 and corrected at cluster level FDR p b 0.05.

ROI analysesTo perform the analysis of regions of interest (ROIs), masks of left

and right V1/V2, V3/V4, and V5 based on a cytoarchitectonic maximumprobability map (Eickhoff, et al., 2005) were generated using SPM

Anatomy Toolbox version 1.8 (http://www.fz-juelich.de/inm/inm-1/spm_anatomy_toolbox). Masks of left and right FG1 and FG2 as de-scribed in Caspers et al. (2013) were used. The mask of VWFA was a10 mm sphere around the classical coordinates (MNI [−42, −57,−12]) (Cohen et al., 2002). We flipped the mask of VWFA to get itscounter-part in right hemisphere (rVWFA). The masks of lateral occipi-tal areas (LO) were based on the centroids of LO1 and LO2 provided byLarson and Heeger (Larsson and Heeger, 2006) and generated withMarsbar (Brett et al., 2002).

For the early retinotopic areas (V1/V2 and V3/V4), we localizedregions corresponding to the horizontal and vertical meridians byasking the participants to go through a localizer run after the mainexperiment. The localizer run included 25 blocks of flashing horizon-tal checkerboard and 25 blocks of flashing vertical checkerboard.Within each hemisphere, ROIs more sensitive to stimuli along thehorizontal meridian (H meridian) in V1/V2 and V3/V4 were deter-mined by selecting the 30 voxels most responsive to horizontalthan to vertical checkerboards. These subject-specific ROIs werethen used to extract response to horizontally presented stimuli. Con-versely, ROIs more sensitive to stimuli along the vertical meridian (Vmeridian) were determined by selecting the 30 most active voxelsshowing the opposite pattern. These subject-specific ROIs werethen used to extract responses to the vertically presented stimuli.For higher visual regions, fixed subject-independent masks wereused, because in those regions the meridian localizer no longer pro-vided systematic distinctions of horizontal and vertical meridianswithin each subject, consistent with previous publications onretinotopy (Engel et al., 1997; Wotawa et al., 2005).

To test the effect of letter status in the ROI analysis, a paired T-testwas applied to each ROI under the horizontal and vertical presentationconditions. To test the frequency effect and the interaction between let-ter status and the other factors, we used a mixed model with partici-pants as random effects and letter status, orientation of presentation,natural-scene frequency, and letter frequency as fixed effects.

Results

Behavioral resultAcross participants, the mean accuracy of the colored item detection

task was 98 % (SE = 0.7 %, range = 88-100%) and the mean RT was468 ms (SE = 9.17 ms, range = 367–535 ms). The participant withthe lowest accuracy (85 %) also yielded the longest RT (624ms). Consid-ering the difference in performance between this participant and thegroup average, this participant was excluded from further analysis.

434 C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

Imaging results

Letter status effect. In the early retinotopic areas, ROI analysis revealedlarger activation for letters than non-letters only in the horizontal pre-sentation condition and only in the left V1/V2 area (t(17) = 2.7, p =.016). In this region, letters induced 8% more activation than non-letters (Fig. 6a).

The corresponding whole-brain SPM analysis revealed an interac-tion between letter status and orientation, namely, larger activationfor letters than non-letters when presented in the horizontal meridianthan in the vertical meridian, at an occipital site corresponding to leftV1/V2 (94% of the whole volume of the cluster)(Fig. 6b and Table 3).

In the higher visual cortex, the ROI analysis revealed larger activa-tion for non-letters in all ROIs except the left FG1 and FG2 (Fig. 7a).This effect was found only for horizontal presented stimuli in left V5

B

A

Horizontal(Letters Vertical(Letters - N

L

Fig. 6. fMRI responses to letter and non-letter stimuli in early retinotopic areas. A: ROI analysisvoxels were selected based on their stronger responses to horizontal than to vertical checkerbthese voxels to letter and non-letter stimuli presented in the same orientation (H or V) in theand non-letters, p b .05). B: Whole-brain search for the predicted interaction between letter stnon-letters for horizontal than for vertical stimuli (N = 18, thresholded at T N 3.12, voxelwise

(t(17) = −3.0, p = .008), LO1 (t(17) = −2.8, p = .014), LO2(t(17) = −3.3, p = .004), VWFA (t(17) = −2.1 p = .046), right V5(t(17) = −2.5, p = .023), LO1 (t(17) = −3.5, p = .002), LO2(t(17) = −4.1, p = .0007), FG1 (t(17) = −3.3, p = .003), FG2(t(17)=−2.9, p= .011), and rVWFA (t(17)=−2.6, p= .018). Largeractivation for letters than non-letters was found only in right V5(t(17) = 2.1, p = .047) for vertically presented stimuli.

A significant Interaction between orientation and letter status wasfound in the left V5 (t(547) = −2.5, p = .01),VWFA (t(547) = −2.2,p = .03), right V5 (t(547) = −2.7, p = .007), LO2 (t(547) = −2.6,p = .009), FG1 (t(547) = −3.0, p = .003), FG2 (t(547) = −2.6, p =.01), and rVWFA (t(547) = −3.0, p = .005).

Consistent with the ROI analyses, results of whole-brain SPM analy-ses also revealed larger activation for non-letters than letters (Table 3,non-letters vs. letters) and an interaction between letter status and

- Non-letters) >on-letters)

within anatomically-defined probabilistic maps for V1/V2 and for V3/V4, subject-specificoards (H) or vice-versa (V) in the localizer run. The graphs show the average response ofindependent line configuration runs (* indicates significance difference between lettersatus and presentation orientation. SPMt map for a greater difference between letters andp b .005, uncorrected; clusterwise p b 0.05, FDR corrected).

Table 3Brain regions showing significant effects of letter status, orientation, natural-scenefrequency, letter frequency, or the interactions between them in Experiment 2.

Cluster Size T X Y Z

horizontal(letters vs. non-letters) N vertical(letters vs. non-letters)543 Occipital Calcrine L 4.10 −9 −93 −8

Lingual L 3.54 −9 −90 −16Mid. L 3.48 −15 −99 10

non-letters vs. letters952 Occipital Mid. L 4.29 −33 −85 3

Inf. L 3.55 −47 −81 −8

horizontal(non-letters vs. letters) N vertical(non-letters vs. letters)577 Occipital Mid. R 3.96 42 −82 1

Inf. R 3.27 35 −82 −5Temporal Mid. R 3.48 51 −69 −2

horizontal(letter frequency) N vertical (letter frequency)419 Occipital Cuneus R 3.74 21 −96 10

Sup. R 3.51 26 −93 19

orientation × letter status × letter frequency1351 Occipital Inf. L 4.81 −21 −100 −7

Mid. L 4.36 −42 −90 −5Inf. L 4.28313 −35 −93 −11

825 Occipital Mid. R 4.26 33 −91 4Mid. R 4.13 38 −88 12Inf. R 3.83193 39 −91 −5

number of junctions1346 Occipital Fusiform L 6.88 −39 −77 −16

Inf. L 4.90 −39 −87 −13885 Inf. R 5.65 38 −79 −17

Inf. R 4.15 44 −75 −10Fusiform R 3.68 36 −69 −14

435C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

orientation in higher visual cortex (Table 3, horizontal(non-letters vs.letters) N vertical(non-letters vs. letters)). This interaction againshowed that the increased activations to non-letters than letters weremainly found with the horizontal stimuli.

Natural-scene frequency and letter frequency effects. The fact that wefailed to observed increasing brain activation for configurations with in-creasingly higher natural-scene frequency, either in SPM analyses or inROI analyses, could result from the fact that after learning to read, the vi-sual system is more sensitive to the frequencies of line configurations intexts rather than in the natural environment. To further test this idea,we examined the effect of the logarithm of letter frequency. Letter fre-quency was extracted from French texts, the subject’s native language,and was weighted by the frequencies of the carrier words, regardless ofcase, as provided by www.LEXIQUE.org (New et al., 2001). Note thatthe correlation coefficient between natural-scene frequency and Frenchletter frequency was positive but non-significant (r = 0.52, p = .26).

When we included letter status, orientation, natural-scene frequen-cy, and letter frequency in the model for the ROI analysis, a three-wayinteraction among letter status, orientation, and letter frequency wasobserved in higher visual cortical areas including left LO1 (t(547) =2.1, p = .03), LO2 (t(547) = 2.4, p = .02), FG2 (t(547) = 2.3, p =.02), VWFA (t(547) = 2.6, p = .008), right LO1 (t(547) = 2.7, p =.008), LO2 (t(547) = 2.6, p = .01), FG2 (t(547) = 2.7, p = .007), andrVWFA (t(547) = 2.7, p = .006). The profile of this triple-interactionwas consistent with an increase in activation with letter frequency,but only for letters and only in the horizontal position (see Fig. 7b). Italso showed that non-letters only had larger activations compared toletters with low letter frequencies, but did not differ from high-letter-frequency ones. There was no effect in early retinotopic areas.

We further examined the letter frequency effect separately for hori-zontally presented letters, horizontally presented non-letters, verticallypresented letters, and vertically presented non-letters. Significant letterfrequency effects were only found in horizontal meridian. For letters,configurations with higher letter frequency elicited larger activation,

while non-letters showed the opposite pattern. The regions showing apositive correlation between letter frequency and brain activation forhorizontally presented letters were left LO2 (t(125) = 3.04, p = .003),left FG1 (t(125) = 3.04, p = .003, left FG2 (t(125) = 3.13, p = .002),VWFA (t(125) = 3.50, p = .0006), right V5 (t(125) = 3.18, p = .002,right LO1 (t(125) = 3.50, p = .0006), right LO2 (t(125) = 2.87, p =.005), right FG1 (t(125) = 2.44, p = .02, right FG2 (t(125) = 3.47, p =.0007), and rVWFA (t(125)=3.80, p= .0002). The regions showingneg-ative correlation between letter frequency and brain activation for hori-zontally presented non-letters were left LO1 (t(125) = −2.45, p =.02), right LO1 (t(125) = −2.36, p = .02), right LO2 (t(125) =−2.37,p = .02), rVWFA (t(125) =−2.00, p = .05).

Consistent with the ROI analyses, voxel-based SPM analyses alsoshowed an interaction between letter frequency and orientation, aswell as a three-way interaction among letter frequency, letter status,and presentation orientation (Table 3).

Negative effect of natural-scene frequency and the role of junction number.As reported above, we did not observe any positive correlations betweenbrain activations andnatural-scene frequency. Although a negative corre-lation was found in the voxel-based analysis in bilateral ventral occipito-temporal cortex and in the ROI analysis in right LO2 (t(547)=−2.3, p=.02; t(547)=2.3, p= .03), given the small number of items used, it couldbe due to confounded factors. One such confound could be the number ofline junctions: the correlation coefficient between natural-scene frequen-cy and number of junctions was -.62 (p = 0.10) (Appendix Table 2). In-deed, increased activation for configurations with more junctions wasobserved in the same regions, at a site plausibly corresponding witharea V4 (V4 covered 61% and 30% of the volume of the cluster in rightand left hemisphere respectively) (Fig. 8 and Table 3), and in a modelwhere both variables were included, the negative correlation betweennatural-scene frequency and brain activation was no longer significant,while the number of junctions effect remained. We therefore went backto Experiment 1 and tested the number of junction effect. The same re-gions showing an effect of the number of junctions in Experiment 2were also detected in Experiment 1 at a lower uncorrected threshold(p b .005 voxelwise, uncorrected) (Fig. 8). We therefore conclude thatthe number of junctions drove this effect. Indeed, the finding of a bilateralventral occipito-temporal effect of the presence of line junctions is con-gruent with prior findings by Szwed et al. (2011).

Discussion

In Experiment 1,we tested thehypothesis that thenatural-scene fre-quency distribution of configurations is reflected in the human visualsystem. We found a positive effect, with bilateral occipital activationsincreasing with the frequencies of line configurations, but we alsofound that this effect wasmost likely due to the fact that many frequentconfigurations also depicted letters of the alphabet, and that thefrequency effect disappeared once letter statuswas controlled for. In Ex-periment 2, we manipulated independently the effects of natural-scenefrequency and letter status. As in Experiment 1, in early retinotopicareas, letters elicited more activation than non-letters. This effect wasreversed in the higher visual cortex. Furthermore, no positive natural-scene frequency effect was found. Instead, a letter frequency effectwas observed in the higher visual cortex. For letters, letter frequencywas positively correlated with brain activations, while for non-letters,a tendency for a negative correlation between letter frequency andbrain activations was found. Experiment 2 also included a novel factor,the orientation of presentation of the stimuli, which were arrayedalong either the horizontal or the vertical meridian. We predicted thatletter status effect would be stronger in brain regions correspondingto the horizontal meridian, which is the location where letter stringsare usually presented during reading Indeed, both the letter status effectand the interaction between letter status and letter frequency weremainly found in regions corresponding to the horizontal meridian.

A B

Left hemisphere Right hemisphere

O rientation Letter frequency

Fig. 7. Letter status effect and its interaction with orientation and letter frequency in higher visual cortex. A: Effects of letter status and presentation orientation. Small brackets indicate asignificant difference between letters and non-letters, while large brackets indicate a significant interaction between stimulus orientation and letter status (p b .05). A larger activation fornon-letters than for letters was found inmany higher-level visual areas, only for horizontally presented stimuli. B: Activations evoked by individual line configurations, sorted as a functionof letter frequency in the written language of the subjects. Again, an influence of letter frequency was only observed for horizontally presented stimuli.

436 C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

Finally, aside from those effects of interest, an increased activation forconfigurationswithmore junctionswas observedmainly in bilateral V4.

Line junctions are thought to be useful visual features of mediumsize and complexity along the hierarchy from simple line segments toentire objects or words. In this respect, our finding that line junctionscause an increased activation in area V4 is consistent with hierarchicalmodels of visual recognition (Dehaene et al., 2005; DiCarlo et al.,2012; Rolls and Stringer, 2006; Serre et al., 2007; Ullman, 2007),which assume that a hierarchy of feature detectors of increasing com-plexity underlies the ventral occipito-temporal “what” pathway. A sim-ilar region was previously reported to respond more strongly to line

drawings where the line junctions were preserved than when theywere deleted (Szwed et al., 2011) (peak around y = −70). Behavioralstudies also demonstrate that the presence of diagnostic line junctionsfacilitates the visual identification of objects and words (Beidermanand Cooper, 1991; Biederman, 1987; Szwed et al., 2011).

Our finding of larger activation to letters than to non-letters in earlyretinotopic areas, however, suggests that physical properties such asfeature complexity and size are not the only factor determining the cor-tical representation of visual features. Rather, the history of perceptualexperience, including literacy, must also be considered. This conclusionfits with studies of perceptual learning, showing that extensive training

Fig. 8. Effect of the number of line junctions in Experiments 1 (left) and 2 (right). A regres-sionmodelwith letter status and number of junctionswas used for Experiment 1 (N=18,thresholded at p b .005, voxelwise p b .005, uncorrected), while a contrast in the ANOVAmodel was used in Experiment 2 (N= 18, thresholded at T N 3.12, voxelwise p b .005, un-corrected; clusterwise p b 0.05, FDR corrected).

437C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

to detect a T shape yields increased activation in V1/V2when this shapeis presented in the trained orientation compared to untrained orienta-tions (Sigman et al., 2005). In the reading domain, a similar early visualeffect was observed when contrasting words versus scrambled words,particularly at an occipital cortical site corresponding to the horizontalmeridian in the left hemisphere (Szwed et al., 2011, 2013). Further,this effect was absent for participants who were not native speakers ofthe tested language (Szwed et al., 2013). These findings, together withours, support the hypothesis of an orientation- and location- specificadaptation in the early retinotopic areas, which is experience depen-dent and probably driven by the need for fast and parallel processing(Gilbert et al., 2001b; Gilbert and Sigman, 2007).

It is worth noting that letter selectivity has been previously observedin higher region of the ventral visual pathway, in tasks that required aninteraction between the perceptual system and higher-order regionswithin the reading network. Using a semantic judgment task, Thesenet al. (2012) compared brain responses to letters, non-letters (falsefont), and real words. They found larger brain responses to letter thannon-letters in the lateral posterior fusiform gyrus. This increased neuralactivity was sustained for an extended duration and was concomitantwith the activation of a broad lexico-semantic processing network.Thus, Thesen et al. (2012) suggest that the selectivity to letters inthis area depends on top-down influences accompanying high levelreading tasks. The recent finding that this area’s response to letterdepends not only on previous experience but also on current contextagain suggests a top-down influence (Grotheer & Kovács, 2014).Conversely, the adoption of a low-level perceptual task may explainwhy our study, like previous fMRI studies, did not show letter selec-tivity in lateral posterior fusiform gyrus (Tagamets and Novick,2000; Vinckier et al., 2007).

Interestingly,we found that the letter status effect reversed in highervisual areas, where there was more activation for non-letters than forletters. Similarly, contrasts between T shapes at untrained orientationversus trained orientation (Sigman et al., 2005), pseudo-letters versusreal letters (Vinckier et al., 2007), and rotated words versus words in anormal orientation (Cohen et al., 2008) all yielded an increased activa-tion in higher ventral occipito-temporal cortex. Those effects mightreflect an on-line top-down influence, such as additional attentionto unfamiliar configurations (Vinckier et al., 2007) or, conversely, de-creased activation to familiar configurations, due to the possibility oftop-downpredictions (Price andDevlin, 2011). Since such top-down in-fluences are known to be context-dependent (Gilbert et al., 2001a; Priceand Devlin, 2011), the horizontal presentation could have offered themost appropriate context for letters and increased this top-down

influence, thus providing a tentative explanation for why such an effectwas only found with the horizontal stimuli in Experiment 2.

We also found a letter frequency effect in higher visual cortex. Con-sistent with this observation, previous studies found a larger activationfor frequent letters and their combinations than for infrequent ones, aneffect which grew from posterior to anterior occipital regions (Binderet al., 2006; Vinckier et al., 2007). We also observed, more surprisingly,a negative correlation between letter frequency and brain activationsfor non-letters. Non-letters with low letter frequencies yielded largeractivations and accounted for the reversed letter status effect in highervisual cortex. This might reflect the fact that high-frequency letters aremore resistant to rotations, thus facilitating their recognition underrotated conditions. Such resistance to rotation could result from neuralrepresentations generalized over broader angles (Ahissar and Hochstein,2004; Folta, 2003).

While the effects of letter status and letter frequencies were salient,across two experiments, the current study did not provide any evidencefor a natural-scene frequency effect. We did find an effect of natural-scene frequency in early retinotopic areas in Experiment 1, but itseemed to be entirely imputable to the presence of letters amongstthemost frequent stimuli, and vanished once this factor was controlledin Experiment 2. Why did reading experience have such a massive im-pact on the visual processing of line configurations, while experiencewith natural scenes seemed to have no impact? Since we scannedstudents, one explanation is that letters have become themost frequentline configuration stimuli in their cultural environment, overriding any(putative) prior effect of natural scenes. Another explanation, not in-compatible with the first one, is the distinction between active and pas-sive perceptual learning. Attention and task requirements have beenshown to deeply influence perceptual learning (Crist et al., 2001; Liet al., 2004, 2008; McManus et al., 2011). In their absence, perceptuallearning is very reduced and occurs only under restricted conditions,e.g. when the unattended stimuli are paired up with attended stimuli(Seitz and Watanabe, 2003) or rewards (Seitz and Watanabe, 2009).Thus, lettersmight have benefited from the active and intensive experi-ence of reading acquisition,while natural scenes are only perceived pas-sively for the most part. The difference and interaction between theneural mechanisms underlying active and passive perceptual learningsare still unclear (Sasaki, Nanez, and Watanabe, 2010; Seitz and Dinse,2007). Future studies on this subject will help to shed further light onour findings.

It is worth noting that although extensive training plays an impor-tant role in shaping early visual areas (Gilbert et al., 2001a; Sigmanet al., 2005), there is clearly a limit on early cortical plasticity. Perceptuallearning effect in early visual cortex has so far beenmostly observed forrelatively simple stimuli such as collinear segments (Zhang and Kourtzi,2010), T shapes (Sigman et al., 2005), moving dots (Watanabe et al.,2002), or gratings (Folta, 2003; Frenkel et al., 2006). Convergingevidence indicates that stimuli as complex as whole words, even afterextensive reading experience, continue to rely on higher visual areassuch as the VWFA (Dehaene and Cohen, 2011; Glezer et al., 2009;Glezer and Riesenhuber, 2013). In the current study, in contrast to theletter status effect in the early retinotopic areas, a bilateral letter fre-quency effect was only found in higher visual cortex. This result is in ac-cordance with the local combination detectors model (Dehaene et al.,2005) and empirical data showing that case- and location-invarianceis only achieved in higher visual cortex (Dehaene et al., 2001, 2004).The complexity of the shapes that can be recognized by neurons in agiven area is likely to be strongly constrained by the underlying neuralcircuitry. For example, it is proposed that the horizontal connectionsbetween pyramidal cells in V1 (Gilbert and Wiesel, 1989; Stettleret al., 2002) enable subsets of neurons to represent complex visualfeatures by integrating information beyond the classical receptivefield (Gilbert et al., 2001b; Li et al., 2006, 2008; McManus et al.,2011). As a consequence, perceptual learning in V1 is likely to beconstrained by the spatial extent of these connections, which

438 C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

extends over about 4 degree of visual space (Stettler et al., 2002),large enough to detect a simple configuration such as a letter, butprobably not an entire word.

In summary, our main finding, across two experiments, is that theearly visual cortex is highly attuned to literacy, to such an extent thatlearned letters induce a stronger activation than matched line configura-tions, especially when presented at the usual horizontal location whereletters usually appear in written texts. Those results reaffirm the impor-tance of literacy acquisition in shaping the human visual system(Dehaene et al., 2010; Pegado and Nakamura, 2014; Szwed et al., 2012).Nevertheless, the present study suffers from several limitations. First, itwould have been useful to obtain a complete subject-specificdelineation of visual areas and retinotopic maps. We did use a within-subject localizer in Experiment 2, but given the time available for scan-ning, we were only able to define ROIs corresponding to horizontal andvertical meridians in early visual areas. Replicating the present resultsand testing their alignment with full retinotopic maps is an importantgoal for future search. Second, this study is also limited by the fact thatonly educated adults were recruited. Because the impact of letters is sostrong, fMRI studies of educated adults are not ideal to properly evaluatethe original hypothesis proposed by Changizi et al. (2006), according to

Appendix Table 1Parameters for each configuration type in Experiment 1.

Configuration Log natural-scenefrequency

Log letterfrequency

Number ofjunctions

Number ofterminals

Numclosu

−1.44 −2.38 1 4 0−0.67 −1.16 1 3 0−0.66 −1.25 1 2 0−4.30 3 2 1−4.30 3 2 1−3.07 3 1 1−2.71 3 2 1−2.30 1 4 0−1.77 2 5 0−1.62 2 4 0−1.59 2 6 0−1.51 −2.54 1 3 0−1.47 −2.05 2 4 0

−1.16 −1.98 2 3 0

−0.94 2 4 0

Appendix A

Appendix Table 2Parameters for each configuration type in Experiment 2.

Configuration Log natural-scenefrequency

Log letterfrequency

Number ofjunctions

Numterm

A −1.12 −2.71 3 2F −1.98 −1.16 2 3H −2.05 −1.47 2 4K −3.52 −2.03 1 4L −1.25 −0.66 1 2T −1.16 −0.67 1 3X −2.38 −1.44 1 4Y −2.54 −1.51 1 3

which the visual system should also be attuned to natural-scene statistics.Future work should endeavor to replicate the present design, searchingfor natural-scene frequency effects in illiterate subjects (Dehaene et al.,2010), in children prior to the acquisition of reading (Monzalvo et al.,2012), or in monkeys without specific symbol training (Brincat andConnor, 2004; Hung et al., 2012; Yau et al., 2012).

Acknowledgments

Wewould like to thank the LBIOM team of the NeuroSpin center fortheir help in subject recruitment and scanning. We gratefully acknowl-edge Alexis Amadon, from NeuroSpin, for his help with high-resolutionfMRI scanning, and Kamil Ugurbil, Essa Yacoub, Steen Moeller, EddieAuerbach and Gordon Junqian Xu, from the Center for MagneticResonance Research, University of Minnesota, for sharing theirpulse sequence and reconstruction algorithms. This research wasfunded by INSERM, CEA, Collège de France, University Paris XI, AgenceNationale de Recherche, and by TAIWAN GRANT NSC 99-2911-I-008-507, NSC 100-2911-I-008-505, NSC 102-2410-H-010-004-MY2, andNSC 102-2911-I-010-507.

ber ofres

Number ofangles

Number ofstrokes

Letterstatus

Convex area(pixel)

RT Accuracy

4 2 1 448 446 96%2 2 1 475 460 94%1 2 1 505 463 97%5 3 0 256 446 99%6 3 0 230 429 98%4 3 0 228 452 94%5 3 0 270 449 99%4 3 0 357 453 96%6 3 0 323 440 99%4 3 0 398 442 99%8 3 0 312 434 99%3 3 1 523 447 95%4 3 1 451 439 98%

3 3 1 379 438 98%

4 3 0 345 442 97%

ber ofinals

Number ofclosures

Number ofangles

Number ofstrokes

Convex area(pixel)

1 5 3 3610 2 3 3880 4 3 3380 3 3 3470 1 2 5200 2 2 5030 4 2 3980 3 3 550

439C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

References

Ahissar, M., Hochstein, S., 2004. The reverse hierarchy theory of visual perceptual learn-ing. Trends Cogn. Sci. 8 (10), 457–464. http://dx.doi.org/10.1016/j.tics.2004.08.011.

Berkes, P., Orbán, G., Lengyel, M., Fiser, J., 2011. Spontaneous cortical activity reveals hall-marks of an optimal internal model of the environment. Science (NewYork, N.Y.) 331(6013), 83–87. http://dx.doi.org/10.1126/science.1195870.

Biederman, I., 1987. Recognition-by-components: a theory of human image understand-ing. Psychol. Rev. 94 (2), 115–147 (Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/3575582).

Beiderman, I., Cooper, E., 1991. Priming contour deleted images. Cogn. Psychol. 23,393–419 Retrieved from http://scholar.google.com/scholar?hl=en&btnG=Search&q=intitle:Priming+contour-deleted+images#8.

Binder, J.J.R., Medler, D. a, Westbury, C.F., Liebenthal, E., Buchanan, L., 2006. Tuning of thehuman left fusiform gyrus to sublexical orthographic structure. NeuroImage 33 (2),739–748 (Retrieved from http://www.sciencedirect.com/science/article/pii/S1053811906007075).

Blakemore, C., Cooper, G., 1970. Development of the brain depends on the visual environ-ment. Nature 228 (5270), 477–478. http://dx.doi.org/10.1038/228477a0.

Brett, M., Anton, J.-L., Valabrègue, R., Poline, J.-B., 2002. Region of interest analysis using anSPM toolbox. 8th International Conference on Functional Mapping of the Human Brain.Sendai, Japan.

Brincat, S.L., Connor, C.E., 2004. Underlying principles of visual shape selectivity in posteriorinferotemporal cortex. Nat. Neurosci. 7 (8), 880–886. http://dx.doi.org/10.1038/nn1278.

Brincat, S.L., Connor, C.E., 2006. Dynamic shape synthesis in posterior inferotemporalcortex. Neuron 49 (1), 17–24. http://dx.doi.org/10.1016/j.neuron.2005.11.026.

Brunswik, E., 1956. Perception and the representative design of psychological experi-ments. vol. 6 (Retrieved from http://www.google.com.tw/books?hl=zh-TW&lr=&id=xTwwQvk6XCUC&pgis=1).

Caspers, J., Zilles, K., Eickhoff, S.B., Schleicher, A., Mohlberg, H., Amunts, K., 2013.Cytoarchitectonical analysis and probabilistic mapping of two extrastriate areasof the human posterior fusiform gyrus. Brain Struct. Funct. 218 (2), 511–526.http://dx.doi.org/10.1007/s00429-012-0411-8.

Changizi, M.A., Zhang, Q., Ye, H., Shimojo, S., 2006. The structures of letters and symbolsthroughout human history are selected to match those found in objects in naturalscenes The, S., Naturalist, A., & May, N. Am. Nat. 167 (5), E117–E139. http://dx.doi.org/10.1086/502806.

Cohen, L., Lehéricy, S., Chochon, F., Lemer, C., 2002. Language‐specific tuning of visual cor-tex? Functional properties of the Visual Word Form Area. Brain 125, 1054–1069 Re-trieved from http://brain.oxfordjournals.org/content/125/5/1054.short.

Cohen, L., Dehaene, S., Vinckier, F., Jobert, A., Montavont, A., 2008. Reading normal and de-graded words: contribution of the dorsal and ventral visual pathways. NeuroImage40 (1), 353–366. http://dx.doi.org/10.1016/j.neuroimage.2007.11.036.

Crist, R., Li,W., Gilbert, C., 2001. Learning to see: experience and attention in primary visu-al cortex. Nat. Neurosci. 4, 519–525 Retrieved from http://www.nature.com/neuro/journal/v4/n5/abs/nn0501_519.html.

Dehaene, S., 2009. Reading in the brain: The new science of how we read. (Retrievedfrom http://books.google.com/books?hl=zh-TW&lr=&id=NlYsTqta7SYC&pgis=1).

Dehaene, S., Cohen, L., 2011. The unique role of the visual word form area in reading.Trends Cogn. Sci. 15 (6), 254–262. http://dx.doi.org/10.1016/j.tics.2011.04.003.

Dehaene, S., Naccache, L., Cohen, L., Bihan, D.L., Mangin, J.F., Poline, J.B., Rivière, D., 2001.Cerebral mechanisms of word masking and unconscious repetition priming. Nat.Neurosci. 4 (7), 752–758. http://dx.doi.org/10.1038/89551.

Dehaene, S., Jobert, A., Naccache, L., Ciuciu, P., Poline, J.B., Le Bihan, D., Cohen, L., 2004.Letter binding and invariant recognition of masked words: behavioral and neuroim-aging evidence. Psychol. Sci. 15 (5), 307–313. http://dx.doi.org/10.1111/j.0956-7976.2004.00674.x.

Dehaene, S., Cohen, L., Sigman, M., Vinckier, F., 2005. The neural code for written words: aproposal. Trends Cogn. Sci. 9 (7), 335–341. http://dx.doi.org/10.1016/j.tics.2005.05.004.

Dehaene, S., Pegado, F., Braga, L.W., Ventura, P., Nunes Filho, G., Jobert, A., …, Cohen, L.,2010. How learning to read changes the cortical networks for vision and language.Science 330 (6009), 1359–1364. http://dx.doi.org/10.1126/science.1194140.

DiCarlo, J.J., Zoccolan, D., Rust, N.C., 2012. How does the brain solve visual object recogni-tion? Neuron 73 (3), 415–434. http://dx.doi.org/10.1016/j.neuron.2012.01.010.

Eickhoff, S., Stephan, K., Mohlberg, H., 2005. A new SPM toolbox for combining probabi-listic cytoarchitectonic maps and functional imaging data. Neuroimage 25 (4),1325–1335. http://dx.doi.org/10.1016/j.neuroimage.2004.12.034.

Engel, S., Glover, G., Wandell, B., 1997. Retinotopic organization in human visual cortexand the spatial precision of functional MRI. Cereb. Cortex 7 (2), 181–192 Retrievedfrom http://cercor.oxfordjournals.org/content/7/2/181.short.

Feinberg, D., Moeller, S., Smith, S., 2010.Multiplexed echo planar imaging for sub−secondwhole brain FMRI and fast diffusion imaging. PloS one 5 (12), e15710 Retrieved fromhttp://dx.plos.org/10.1371/journal.pone.0015710.

Folta, K., 2003. Neural fine tuning during Vernier acuity training? Vis. Res. 43 (10),1177–1185. http://dx.doi.org/10.1016/S0042-6989(03)00041-5.

Frenkel, M.Y., Sawtell, N.B., Diogo, A.C.M., Yoon, B., Neve, R.L., Bear, M.F., 2006. Instructiveeffect of visual experience in mouse visual cortex. Neuron 51 (3), 339–349. http://dx.doi.org/10.1016/j.neuron.2006.06.026.

Furmanski, C.S., Engel, S.A., 2000. An oblique effect in human primary visual cortex. Nat.Neurosci. 3 (6), 535–536. http://dx.doi.org/10.1038/75702.

Geisler, W.S., 2008. Visual perception and the statistical properties of natural scenes. Annu.Rev. Psychol. 59, 167–192. http://dx.doi.org/10.1146/annurev.psych.58.110405.085632.

Gilbert, C.D., Sigman, M., 2007. Brain states: top-down influences in sensory processing.Neuron 54 (5), 677–696. http://dx.doi.org/10.1016/j.neuron.2007.05.019.

Gilbert, C.D., Wiesel, T.N., 1989. Columnar specificity of intrinsic horizontal andcorticocortical connections in cat visual cortex. J. Neurosci. Off. J. Soc. Neurosci. 9(7), 2432–2442 (Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/2746337).

Gilbert, C.D., Sigman, M., Crist, R., 2001a. The neural basis of perceptual learning. Neuron31 (5), 681–697 (Retrieved from http://www.sciencedirect.com/science/article/pii/S089662730100424X).

Gilbert, C.D., Sigman, M., Crist, R.R.E., 2001b. The neural basis of perceptual learning. Neu-ron 31 (5), 681–697 (Retrieved from http://www.sciencedirect.com/science/article/pii/S089662730100424X).

Girshick, A.R., Landy, M.S., Simoncelli, E.P., 2011. Cardinal rules: visual orientation percep-tion reflects knowledge of environmental statistics. Nat. Neurosci. 14 (7), 926–932.http://dx.doi.org/10.1038/nn.2831.

Glezer, L.S., Riesenhuber, M., 2013. Individual variability in location impacts orthographicselectivity in the “visual word form area”. J. Neurosci. Off. J. Soc. Neurosci. 33 (27),11221–11226. http://dx.doi.org/10.1523/JNEUROSCI.5002-12.2013.

Glezer, L.S., Jiang, X., Riesenhuber, M., 2009. Evidence for highly selective neuronal tuningto whole words in the “visual word form area”. Neuron 62 (2), 199–204. http://dx.doi.org/10.1016/j.neuron.2009.03.017.

Grotheer, M., Kovács, G., 2014. Repetition probability effects depend on prior experiences.J. Neurosci. 34 (19), 6640–6646. http://dx.doi.org/10.1523/JNEUROSCI.5326-13.2014.

Harvey, B.M., Dumoulin, S.O., 2011. The relationship between cortical magnification factorand population receptive field size in human visual cortex: constancies in cortical ar-chitecture. J. Neurosci. Off. J. Soc. Neurosci. 31 (38), 13604–13612. http://dx.doi.org/10.1523/JNEUROSCI.2572-11.2011.

Held, R., Hein, A., 1963. Movement−produced stimulation in the development of visuallyguided behavior. J. Comp. Physiol. Psychol. 56 (5), 872–876 Retrieved from http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.211.8253.

Hess, R.F., Hayes, A., Field, D.J., 2003. Contour integration and cortical processing.J. Physiol. Paris 97 (2–3), 105–119. http://dx.doi.org/10.1016/j.jphysparis.2003.09.013.

Howe, C.Q., Purves, D., 2002. Range image statistics can explain the anomalous perceptionof length. Proc. Natl. Acad. Sci. U. S. A. 99 (20), 13184–13188. http://dx.doi.org/10.1073/pnas.162474299.

Hung, C.-C., Carlson, E.T., Connor, C.E., 2012. Medial axis shape coding in macaqueinferotemporal cortex. Neuron 74 (6), 1099–1113. http://dx.doi.org/10.1016/j.neuron.2012.04.029.

Kersten, D., Mamassian, P., Yuille, A., 2004. Object perception as Bayesian inference. Annu.Rev. Psychol. 55, 271–304. http://dx.doi.org/10.1146/annurev.psych.55.090902.142005.

Knill, D.C., Pouget, A., 2004. The Bayesian brain: the role of uncertainty in neural codingand computation. Trends Neurosci. 27 (12), 712–719. http://dx.doi.org/10.1016/j.tins.2004.10.007.

Larsson, J., Heeger, D.J., 2006. Two retinotopic visual areas in human lateral occipital cortex.J. Neurosci. 26 (51), 13128–13142. http://dx.doi.org/10.1523/jneurosci.1657-06.2006.

Li,W., Piëch, V., Gilbert, C.D., 2004. Perceptual learning and top-down influences in primaryvisual cortex. Nat. Neurosci. 7 (6), 651–657 (Retrieved from http://www.nature.com/neuro/journal/v7/n6/abs/nn1255.html).

Li, W., Piëch, V., Gilbert, C.D., 2006. Contour saliency in primary visual cortex. Neuron 50(6), 951–962. http://dx.doi.org/10.1016/j.neuron.2006.04.035.

Li, W., Piëch, V., Gilbert, C.D., 2008. Learning to link visual contours. Neuron 57 (3),442–451. http://dx.doi.org/10.1016/j.neuron.2007.12.011.

Long, F., Purves, D., 2003. Natural scene statistics as the universal basis of color contexteffects. Proc. Natl. Acad. Sci. U. S. A. 100 (25), 15190–15193. http://dx.doi.org/10.1073/pnas.2036361100.

Maldjian, J.a., Laurienti, P.J., Kraft, R.a., Burdette, J.H., 2003. An automatedmethod for neu-roanatomic and cytoarchitectonic atlas-based interrogation of fMRI data sets.NeuroImage 19 (3), 1233–1239. http://dx.doi.org/10.1016/S1053-8119(03)00169-1.

McManus, J.N.J., Li, W., Gilbert, C.D., 2011. Adaptive shape processing in primary visualcortex. Proc. Natl. Acad. Sci. U. S. A. 108 (24), 9739–9746. http://dx.doi.org/10.1073/pnas.1105855108.

Moeller, S., Yacoub, E., Olman, C.A., Auerbach, E., Strupp, J., Harel, N., Uğurbil, K., 2010.Multiband multislice GE-EPI at 7 tesla, with 16-fold acceleration using partial parallelimaging with application to high spatial and temporal whole-brain fMRI. Magn.Reson. Med. 63 (5), 1144–1153. http://dx.doi.org/10.1002/mrm.22361.

Monzalvo, K., Fluss, J., Billard, C., Dehaene, S., Dehaene-Lambertz, G., 2012. Cortical net-works for vision and language in dyslexic and normal children of variable socio-economic status. NeuroImage 61 (1), 258–274. http://dx.doi.org/10.1016/j.neuroimage.2012.02.035.

Nazir, T. a, Ben-Boutayab, N., Decoppet, N., Deutsch, A., Frost, R., 2004. Reading habits, per-ceptual learning, and recognition of printed words. Brain Lang. 88 (3), 294–311.http://dx.doi.org/10.1016/S0093-934X(03)00168-8.

New, B., Pallier, C., Ferrand, L., Matos, R., 2001. Une base de données lexicales du françaiscontemporain sur internet: LEXIQUE. Annee Psychol. 101 (3), 447–462 (Retrievedfrom http://dialnet.unirioja.es/servlet/articulo?codigo=3228917&info=resumen&idioma=ENG).

Pegado, F., Nakamura, K., 2014. Literacy breaks mirror invariance for visual stimuli: a be-havioral study with adult illiterates. J. Exp. Psychol. 143 (2), 887–894 Retrieved fromhttp://psycnet.apa.org/journals/xge/143/2/887/.

Price, C.J., Devlin, J.T., 2011. The interactive account of ventral occipitotemporal contribu-tions to reading. Trends Cogn. Sci. 15 (6), 246–253. http://dx.doi.org/10.1016/j.tics.2011.04.001.

Rolls, E.T., Stringer, S.M., 2006. Invariant visual object recognition: a model, with lightinginvariance. J. Physiol. Paris 100 (1–3), 43–62. http://dx.doi.org/10.1016/j.jphysparis.2006.09.004.

Sasaki, Y., Nanez, J.E., Watanabe, T., 2010. Advances in visual perceptual learning and plas-ticity. Nat. Rev. Neurosci. 11 (1), 53–60. http://dx.doi.org/10.1038/nrn2737.

440 C.H.C. Chang et al. / NeuroImage 120 (2015) 428–440

Seitz, A.R., Watanabe, T., 2003. Psychophysics: Is subliminal learning really passive?Nature 422 (6927), 36. http://dx.doi.org/10.1038/422036a.

Seitz, A.R., Dinse, H.R., 2007. A common framework for perceptual learning. Curr. Opin.Neurobiol. 17 (2), 148–153. http://dx.doi.org/10.1016/j.conb.2007.02.004.

Seitz, A.R., Watanabe, T., 2009. The phenomenon of task-irrelevant perceptual learning.Vis. Res. 49 (21), 2604–2610. http://dx.doi.org/10.1016/j.visres.2009.08.003.

Serre, T., Oliva, A., Poggio, T., 2007. A feedforward architecture accounts for rapid catego-rization. Proc. Natl. Acad. Sci. U. S. A. 104 (15), 6424–6429. http://dx.doi.org/10.1073/pnas.0700622104.

Shepard, R.N., 2002. Perceptual-cognitive universals as reflections of the world. Behav.Brain Sci. 24 (04), 581–601. http://dx.doi.org/10.1017/S0140525X01000012.

Sigman, M., Pan, H., Yang, Y., Stern, E., Silbersweig, D., Gilbert, C.D., 2005. Top-down reor-ganization of activity in the visual pathway after learning a shape identification task.Neuron 46 (5), 823–835. http://dx.doi.org/10.1016/j.neuron.2005.05.014.

Simoncelli, E., Olshausen, B., 2001. Natural image statistics and neural representation.Annu. Rev. Neurosci. 24, 1193–1216 Retrieved from http://www.annualreviews.org/doi/pdf/10.1146/annurev.neuro.24.1.1193.

Stettler, D.D., Das, A., Bennett, J., Gilbert, C.D., 2002. Lateral connectivity and contextualinteractions in macaque primary visual cortex. Neuron 36 (4), 739–750 (Retrievedfrom http://www.sciencedirect.com/science/article/pii/S0896627302010292).

Szwed, M., Dehaene, S., Kleinschmidt, A., Eger, E., Valabrègue, R., Amadon, A., Cohen, L.,2011. Specialization for written words over objects in the visual cortex. NeuroImage56 (1), 330–344. http://dx.doi.org/10.1016/j.neuroimage.2011.01.073.

Szwed, M., Ventura, P., Querido, L., Cohen, L., Dehaene, S., 2012. Reading acquisitionenhances an early visual process of contour integration. Dev. Sci. 15 (1), 139–149.http://dx.doi.org/10.1111/j.1467-7687.2011.01102.x.

Szwed, M., Qiao, E., Jobert, A., Dehaene, S., Cohen, L., 2013. Effects of literacy in early visualand occipitotemporal areas of Chinese and French readers. 1–33 (Retrieved fromhttp://www.mitpressjournals.org/doi/abs/10.1162/jocn_a_00499).

Szwed, M., Qiao, E., Jobert, A., Dehaene, S., Cohen, L., 2014. Effects of literacy in earlyvisual and occipitotemporal areas of Chinese and French readers. (Retrieved fromhttp://www.mitpressjournals.org/doi/abs/10.1162/jocn_a_00499#.VCgW8RaTBdQ).

Tagamets, M., Novick, J., 2000. A parametric approach to orthographic processing in thebrain: an fMRI study. J. Cogn. Neurosci. 12 (2), 281–297 Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/10771412.

Tanaka, K., 2003. Columns for complex visual object features in the inferotemporal cortex:clustering of cells with similar but slightly different stimulus selectivities. Cereb. Cortex13 (1), 90–99. http://dx.doi.org/10.1093/cercor/13.1.90.

Thesen, T., McDonald, C.R.C., Carlson, C., Doyle, W., Cash, S., Sherfey, J., Felsovalyi, O.,Girard, H., Barr, W., Devinsky, O., Kuzniecky, R., Halgren, E., 2012. Sequential then in-teractive processing of letters and words in the left fusiform gyrus. Nat. Commun. 3,1284. http://dx.doi.org/10.1038/ncomms2220.

Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N.,Mazoyer, B., Joliot, M., 2002. Automated Anatomical Labeling of Activations in SPMUsing a Macroscopic Anatomical Parcellation of the MNI MRI Single−Subject Brain.Neuroimage 15 (1), 273–289. http://dx.doi.org/10.1006/nimg.2001.0978.

Ullman, S., 2007. Object recognition and segmentation by a fragment-based hierarchy.Trends Cogn. Sci. 11 (2), 58–64. http://dx.doi.org/10.1016/j.tics.2006.11.009.

Vinckier, F., Dehaene, S., Jobert, A., Dubus, J.P., Sigman, M., Cohen, L., 2007. Hierarchicalcoding of letter strings in the ventral stream: dissecting the inner organization ofthe visual word-form system. Neuron 55 (1), 143–156. http://dx.doi.org/10.1016/j.neuron.2007.05.031.

Watanabe, T., Náñez, J.E., Koyama, S., Mukai, I., Liederman, J., Sasaki, Y., 2002. Greater plas-ticity in lower-level than higher-level visual motion processing in a passive perceptu-al learning task. Nat. Neurosci. 5 (10), 1003–1009. http://dx.doi.org/10.1038/nn915.

Wotawa, N., Thirion, B., Castet, E., Anton, J.−L., Faugeras, O., 2005. Human retinotopicmapping using fMRI. Technical Report 5472, INRIA Sophia−Antipolis.

Xu, J., Moeller, S., Auerbach, E., Strupp, J., 2013. Evaluation of slice accelerations usingmul-tiband echo planar imaging at 3T. Neuroimage 93, 991–1001 Retrieved from http://www.sciencedirect.com/science/article/pii/S1053811913008240.

Yau, J., Pasupathy, A., Brincat, S., Connor, C., 2012. Curvature processing dynamics inmacaque area V4. Cereb. Cortex 23 (1), 198–209. http://dx.doi.org/10.1093/cercor/bhs004.

Zhang, J., Kourtzi, Z., 2010. Learning-dependent plasticity with and without training in thehuman brain. Proc. Natl. Acad. Sci. U. S. A. 107 (30), 13503–13508. http://dx.doi.org/10.1073/pnas.1002506107.